- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
CFGP replaces per-task prompt optimization with a prompt computed from support-set class prototypes and clustering assignments. The prompt modifies the input features, which are then re-encoded by a frozen GNN. The method supports both node and graph classification. Its main appeal is the reduced adaptation cost, with competitive results on several of the eight evaluated datasets.
Critical strengths
Removing the backward pass makes adaptation substantially cheaper. The construction is straightforward and works for both task types without training a separate prompt module.
Critical weaknesses
-
My main concern is the use of evaluation tasks for configuration selection. Appendix A specifies a search over 162 configurations per dataset, evaluated on the same tasks used to report the final results. This is test-set tuning, even though the prediction procedure itself does not read test labels. The shared-configuration result is encouraging, but the paper needs to explain how that configuration was chosen and report results on tasks excluded from selection. The timing comparison should also account separately for configuration search.
-
The baseline changes in Appendix E make several comparisons difficult to interpret. GPF, GPF+, All-in-One, and SUPT are applied to cached encoder outputs, changing their intended interaction with the GNN. These results do not establish an advantage over faithful implementations of those methods. I would also expect comparisons with simple training-free alternatives, particularly label propagation and transductive prototype refinement.
Questions for authors
See weaknesses